fix(pain): propagate agent_id through pain reaction reward chain

[dennys246]

Summary

Fixes the silent failure mode that left _reward_bias empty across every cradle / damage / pain-bearing sim despite pain firing constantly. Root cause: pain reactions emerged with ReactionContext.agent_id=None, the reward-distributor subscriber early-returned on every one, and distribute() was never called.

Investigation

Added a structured trace gated by MAXIM_PAIN_CHAIN_TRACE=1 that emits one event per chain transition:

pain_chain.bus_publish       (PainBus.publish entry)
pain_chain.reaction_emitted  (after compat conversion)
pain_chain.reward_subscriber (subscriber receives reaction)
pain_chain.distribute_returned (after distributor.distribute call)

First cradle run with trace (no fix):

Event	Count	Note
bus_publish	11	pain emits — works
reaction_emitted	11	compat converts
reward_subscriber	19	receives, but will_distribute: false on every one
distribute_returned	0	chain breaks here

bus_publish events showed ctx_has_agent_id: false for every pain — body publishers weren't propagating agent_id, and even if they had, pain_signal_to_reaction wouldn't have extracted it into ReactionContext.

Fix (three parts)

1. Embodiment.__init__ takes agent_id: str = "" and stores it. Both _publish_pain and _publish_drive_pain include it in PainSignal.context. runtime/bootstrap.py passes agent_id=agent_id when constructing the embodiment.
2. simulation/tools.py::DamageComponentTool (the third pain publisher — fires from orchestrator-side damage calls including reflex_fire) reads agent_id from self._embodiment and adds it to its inline-constructed PainSignal.context.
3. reactions/compat.py::pain_signal_to_reaction extracts signal.context["agent_id"] and passes it to ReactionContext(agent_id=...). Empty string normalises to None so the subscriber's documented agent_id is None early-return fires correctly for foundry / scene-entity embodiments that aren't learning subjects.

Verification — third cradle run with all fixes:

Event	Count
bus_publish	17
reaction_emitted	17
reward_subscriber	21 (will_distribute: true, agent_id="sim_aut")
distribute_returned	12 (11 of 12 credited 1+ nodes)

The chain is fully wired. goal_reward_bias accumulates (cradle: -0.16) confirming the distributor reaches its terminal stage.

On _reward_bias still being empty

Verified: per the existing CLAUDE.md invariant, "_reward_bias clamps to [0, max_reward_bias] — Negative rewards (pain) produce 0.0 bias. Pain avoidance is handled by valence annotation on edges, not by reward bias." Negative pain rewards clamp to 0; the per-tick decay then prunes 0.0 entries. This is intended. Pain-only sims produce no persistent reward_bias state. Positive Reactions on the bus would now land correctly — that part of the chain was the dormancy this PR closes.

Trace infrastructure

The MAXIM_PAIN_CHAIN_TRACE=1 env gate + pain_chain_trace() helper in proprioception/pain_bus.py ship in this commit. Future debugging of the chain doesn't need to re-instrument it.

Test plan

• Full fast suite: 6357 passed (was 6336; +3 new tests)
• ruff check + ruff format on every touched file
• New tests pin the agent_id propagation contract:
  • test_publish_propagates_agent_id_into_reaction_context — signal context → ReactionContext
  • test_publish_with_empty_agent_id_normalises_to_none — empty string normalisation
  • test_embodiment_includes_agent_id_in_pain_context — Embodiment writes agent_id into pain
• End-to-end cradle sim with trace events shows the chain transitioning cleanly through all four stages

🤖 Generated with Claude Code